The dataset contains 11 folders in total.

2dot:
Three folders, one with 500 samples, one with 1000, one with 5000. All are sampled with replacement.
In the article, the 5000 one is used for training and 1000 used for validation during training.

When used only for evaluation (when training on MNIST-bg), the 500 one is used.

classification_2dot__bg0_5classes_500seqs_20_per_seq
classification_2dot__bg0_5classes_1000seqs_20_per_seq
classification_2dot__bg0_5classes_5000seqs_20_per_seq

5dot:
Two folders, one with 1000 samples and one with 500 samples. Only the one with 500 samples is used in our paper (for unseen evaluation).
Folder names:
classification_5dot__bg0_5classes_500seqs_20_per_seq
classification_5dot__bg0_5classes_1000seqs_20_per_seq

MNIST:
Similar to 5Dot, two folders, one with 1000 samples and one with 500 samples. Only the one with 500 samples is used in our paper (for unseen evaluation). The digits are always taken from the MNIST test set.

(Note that we do not train on 5Dot or MNIST in the paper).
Folder names:
classification_14mnist_test_bg0_5classes_500seqs_20_per_seq
classification_14mnist_test_bg0_5classes_1000seqs_20_per_seq


MNIST-bg:
Four folders. Two with 1000 and 500 samples, containing MNIST training digits, intended for training (we used 5000 samples in our paper).
Folder names:
classification_14mnist_train_bg1_5classes_1000seqs_20_per_seq
classification_14mnist_train_bg1_5classes_5000seqs_20_per_seq

Two with 500 and 1000 samples, intended for evaluation only, containing digits only from the MNIST test set.
Folder names:
classification_14mnist_test_bg1_5classes_500seqs_20_per_seq
classification_14mnist_test_bg1_5classes_1000seqs_20_per_seq

*******************************************************************************************************

All sequences contain 20 frames. To read the right sequence, take the sequence ID from the labels.csv file within one folder, multiply it by 20 and read those 20 subsequent frames starting with sequence_ID*20.jpg. Nevertheless, there is code in the repository https://github.com/sofiabroome/temporal-shape-dataset  showing how to read the sequences.

In each folder, there are example gifs for 100 sequences, for inspection.
